Google combined tools example

This notebook features a more advanced usage of Agent Tools, using the Google Calendar, Mail and Search integrations as well as the Load and Search Meta Tool to fufill a more complicated set of tasks for the user.

Setup the Tools

First we will import OpenAI and setup the Agent:

python

import os

os.environ["OPENAI_API_KEY"] = "sk-api-key"

from llama_index.core.agent.workflow import FunctionAgent
from llama_index.llms.openai import OpenAI

Now we can import the Google Tools we are going to use. See the README for the respective tools to get started with authentication.

python

from llama_index.tools.gmail.base import GmailToolSpec
from llama_index.tools.google_calendar.base import GoogleCalendarToolSpec
from llama_index.tools.google_search.base import GoogleSearchToolSpec

gmail_tools = GmailToolSpec().to_tool_list()
gcal_tools = GoogleCalendarToolSpec().to_tool_list()
gsearch_tools = GoogleSearchToolSpec(key="api-key", engine="engine").to_tool_list()

Let's take a look at all of the tools we have available from the 3 tool specs we initialized:

python

for tool in [*gmail_tools, *gcal_tools, *gsearch_tools]:
    print(tool.metadata.name)
    print(tool.metadata.description)

We have to be conscious of the models context length when using these tools as if we are not careful the response can easily be larger than the token limit. In particular, the load_data function for emails returns large payloads, as well as google search. In this example I will wrap those two tools in the Load and Search Meta tool:

python

from llama_index.tools.tool_spec.load_and_search.base import LoadAndSearchToolSpec

print("Wrapping " + gsearch_tools[0].metadata.name)
gsearch_load_and_search_tools = LoadAndSearchToolSpec.from_defaults(
    gsearch_tools[0],
).to_tool_list()

print("Wrapping gmail " + gmail_tools[0].metadata.name)
gmail_load_and_search_tools = LoadAndSearchToolSpec.from_defaults(
    gmail_tools[0],
).to_tool_list()

print("Wrapping google calendar " + gcal_tools[0].metadata.name)
gcal_load_and_search_tools = LoadAndSearchToolSpec.from_defaults(
    gcal_tools[0],
).to_tool_list()

Notice we are only wrapping individual tools out of the tool list. Lets combine the all the tools together into a combined list:

python

all_tools = [
    *gsearch_load_and_search_tools,
    *gmail_load_and_search_tools,
    *gcal_load_and_search_tools,
    *gcal_tools[1::],
    *gmail_tools[1::],
    *gsearch_tools[1::],
]

Now the tools are ready to pass to the agent:

python

agent = FunctionAgent(tools=all_tools, llm=OpenAI(model="gpt-4.1"))

Interacting with the Agent

We are now ready to interact with the Agent and test the calendar, email and search capabilities! Let's try out the search first:

python

await agent.run(
    "search google and find the email address for a dentist in toronto near bloor and"
    " dufferin"
)

Summary

We were able to use the google search, email and calendar tools to find a dentist and request an appointment, making sure any calendar conflicts were avoided. This notebook should prove useful for experimenting with more complicated agents, combining together the tools available in LlamaHub to create more complicated workflows the agent can execute.